AI agents for finance: what they can do and where they need your approval
What does AI in finance look like today?
Take a question one of our customers had to answer for their CEO, who wanted to hire two more account managers from next quarter: can we afford them? The answer depends on the runway after two more salaries, the cash position, the sales team's spending against its budget and the customer invoices that are still open. In most companies, that means exports from the bank, the invoicing tool and the accounting tool, combined in a spreadsheet. One founder told us they spend one to two days a month on what they called a monster Excel, only to keep their projections up to date.
AI should make questions like this easy, but its use in finance has stalled. In Gartner's 2025 AI in Finance Survey of 183 CFOs and senior finance leaders, 59 percent said their finance function uses AI, barely up from 58 percent the year before, after a jump from 37 percent in 2023. In the meantime, many teams build their own workarounds, like the company we spoke to whose interim bookkeeper is building a posting matrix so that bookings are suggested automatically. Another prepares its accruals with Python scripts written by junior staff, and a third is experimenting with AI tools on its own.
Why does AI in finance so often stay a side project?
Because an AI is only as good as the data it can read, and in most finance setups that data is scattered across tools, which goes wrong in three common ways.
The first is AI reading an export, because a general chatbot does not know your books. It summarises whatever you paste into it, without knowing whether the export was complete, whether credit notes were included or whether anything has been booked since, and it gives the answer with the same confidence either way. Paste a different export next week and you can get a different answer for the same month, with no record of how either one was produced.
The second is AI inside a single tool: the AI in your expense management tool sees card transactions, the one in your invoicing tool sees invoices, and the one in your CRM sees deals. A question about margin needs the revenue from your invoices and the costs from your expense management tool and payroll, so the answer still ends up in a spreadsheet where someone combines the data by hand.
The third is AI reading books that are out of date. If payments are matched to invoices only once a month, an answer on the 20th is based on a ledger that is three weeks behind, and if cost centres are only corrected at month-end, a question about spending by team returns the uncorrected figures.
What can AI do in a finance team?
Once bookings, documents and bank transactions sit in one ledger that is kept current, AI can take on a large part of the routine work. It can read each bank transaction, invoice and receipt and suggest the matching document, the account and the cost centre, so that your team reviews suggestions instead of keying in bookings. It can code supplier invoices, check them against the purchase order and goods receipt in a three-way match and send them to the right approver.
It can also answer questions about your numbers, such as how much cash the company has and how long the runway is at the current burn rate. The same goes for plan versus actual by cost centre, customer invoices that are overdue and supplier invoices that are due this week. And it can turn those figures into the reports your board and investors ask for, as a presentation in your company's branding or as an Excel file in the structure your investors expect.
What should AI in finance not do on its own?
Above all, it should not post bookings without a person approving them. German bookkeeping rules require every booking to be traceable to its document, and a posted entry has to be corrected by reversing it rather than changing it. An AI that posts on its own therefore creates corrections that someone has to find and undo later, while a suggestion with a confidence score that your team approves keeps both the speed and the control.
Every answer should also show what it is made of, meaning the invoices, payments and bookings that add up to a figure, so you can check it before it goes into a decision or a board report. And the AI has to follow the same access rights as the rest of your finance system, so that salaries and individual expense claims only appear to people who can already see them in the books.
How is AI used in Agent F?
In Agent F, AI works at two levels, and the first is how the ERP is built. Through our AI-native ERP Factory, you describe your requirements, your business logic and context, the tools you already use and the problems you want to solve, and Agent F uses AI to configure the ERP around your company, with the modules, workflows and automations it needs. You do not have to fit your company into a standard template or pay for months of custom development.
The second level is how the ERP runs: your bank accounts, expense management tool, invoicing tool and payroll and HR tool are connected, and every transaction is reconciled as it comes in, with the booking suggested and a confidence score for your team to approve. Going back to our example, when your CEO asks whether you can afford the two new account managers, you prompt it into Agent F: "Can we afford two more account managers from next quarter?" The answer shows the runway with the two extra salaries, the cash position and the sales team's spending against its budget. All of it comes from the same live books, with the bookings that make up each figure and only the data your access rights allow.
What would you ask your books first?
Learn how Agent F works, and book a demo at agent-f.ai/demo to see it run on your numbers.